Forum > GPU crunching
V10/11 of modified SETI MB CUDA + opt AP package for full multi-GPU+CPU use
Raistmer:
--- Quote from: RottenMutt on 22 Mar 2009, 10:44:34 am ---I'm running "stock" Boinc 6.6.17-64bit on my lappy, no optimized application installed yet. I immediately noticed that it will keep the GPU fed with work units, it also labels the Enhanced 6.08 work units as "(cuda)". It appears to me the latest Boinc runs cuda applications and fetches cuda applications as you would expect, woot!
I noticed in the c:/program data/.../setiathome.berkeley.edu folder there is not an app_info.xml file. If a install Raistmer's Opt V10a Pack along with a appropriate cc_config.xml file (or none) I end up running one more AP work unit then the number logical CPU's, rather then placing a 6.08 cuda work unit on the GPU.
So how can one use Raistmer's Opt V10a Pack and get it to behave the same as stock??? I think what is in the app_info file needs to be updated for the latest boinc.
what would keep me from renaming Raistmer's executables the same a the stock executable and copy the dll files? i really don't want to do that, they could easily be over written by Bionc.
--- End quote ---
AFAIK you need app_info in all cases when you want to use opt apps. No mater what BOINC version do you use or do you use CPU only or CPU+GPU.
Working w/o app_info will enable only stock apps on your host.
RottenMutt:
what is the benefit of your optimized cuda over stock cuda???
I'm going to try MarkJ's app_info, just wondering if it is worth while to insert your cuda app over the stock.
this should finally get my gpu loaded.
Raistmer:
--- Quote from: RottenMutt on 22 Mar 2009, 01:54:58 pm ---what is the benefit of your optimized cuda over stock cuda???
--- End quote ---
1) VLAR-kill mod
2) correct priority settings
3) elapsed time information availability
[4) affinity binding and ignoring of BOINC suspend order]
Last position is relative advantage, not in all cases it's good.
In coming V11 (at least CUDA app if not team pack) will be reduction of initial CPU time.
But I spoke not only about CUDA app itself. If you will not use app_info.xml you will not be able to use opt AP (~5x performnce loss) and opt AK_v8 (~1,5-2 performance loss).
All opt apps need appropriate app_info to be used.
RottenMutt:
ok, i plugged in your cuda app. i'm running all your apps on Boinc 6.6.17 using MarkJ's app_info.xml, i replaced the stock executables with yours in the app_info file and everything works. i have not had a chance to see if the gpu count option works yet at the bottom of the app_info file.
see the following thread for MarkJ's app_info instructions.
http://setiathome.berkeley.edu/forum_thread.php?id=52589
edit: i have not had to modify the gpu count at the bottom of MarkJ's app_info file on machines with more then one video card.
chronek:
I wonder if could be 2 quenies , one for ap , sec for gpu - all mb go for gpu , and we could set cpu cores for quenie ap , and units gpu with quenie mb , and if ap quenie would be empty they could fetch from sec quenie mb - that could solve problems with gpu idle - and cpu idle , and for cpu - eliminate situation when work at 3 wu on 2 cores
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version